Automate back-office work with approval steps
Agents that call your systems to complete multi-step back-office work, with human approval at the risky steps and a full record of every action.
Overview
An agent that reads a request, looks up records, drafts a reply and updates a system can remove hours of copy and paste. It can also make confident mistakes at speed. The design question is where a person must approve before anything changes.
We build tool-using agents with narrow permissions, explicit approval steps and retries that do not double-post. Where a system has no API, we use RPA for that step. Every run is logged so you can see what the agent did and why.
Signs you need this
- Staff copy the same fields between three systems every day
- An agent prototype works most of the time but nobody trusts it with real records
- You cannot explain afterwards why an automated step did what it did
- Government or supplier portals force manual data entry
What you get
- Process map
- The target workflow broken into steps, each marked as automatic, agent-assisted or human-only, with the reason.
- Agent and tools
- An agent with a small set of typed tools, each mapped to one API call or RPA action with least-privilege credentials.
- Approval interface
- Review screens in Slack, Teams or a web app where a person approves, edits or rejects proposed actions.
- Orchestration
- Durable workflows with retries, timeouts and idempotent steps so a failed run resumes without duplicate actions.
- Run history and audit trail
- Each run stored with inputs, model reasoning summary, tool calls, approvals and final result.
How the work runs
- 1
Process study
1 to 2 weeks
We watch the work being done, collect examples and agree which steps need human approval.
- 2
Build first workflow
3 to 5 weeks
We build the agent, tools, approvals and orchestration for one workflow, tested against past cases.
- 3
Shadow run
2 to 3 weeks
The agent proposes actions alongside staff without changing anything, so we can compare results.
- 4
Go live and handover
1 to 2 weeks
We switch on approved actions, monitor the first weeks and hand over runbooks and dashboards.
Timings depend on the size of your estate. The assessment fixes them before any build starts.
What you can check afterwards
- No record changes in a target system without a logged approval where one is required
- Any run can be replayed from its audit trail to see each tool call and decision
- A run interrupted midway resumes without creating duplicate records
- Shadow-run results are documented next to the staff decisions they were compared with
Tools we work with
- LangGraph
- Temporal
- n8n
- UiPath
- Playwright
- Python
- PostgreSQL
- Slack API
- Microsoft Teams
- OpenTelemetry
Questions buyers ask
- What if the agent gets it wrong?
- Risky actions wait for a person to approve them, and the agent's credentials only allow the actions it needs. Mistakes are caught at approval or limited in scope.
- Does our data go to an external AI provider?
- That depends on the model you choose. We can use a self-hosted model or a provider endpoint in your region, and redact personal fields before any call.
- What access do you need?
- Test accounts in each system the workflow touches, API credentials where they exist, and a few hours with the people who do the work today.
- Who maintains it afterwards?
- Your team, with our runbooks. Portals and APIs change, so we show you how to update a tool and re-run the test cases.
Talk to us about AI agents and workflow automation.
Tell us what you run and where. We reply within one business day with questions or a quote.